def generate_code_for_sched_index(codegen_state, sched_index):
kernel = codegen_state.kernel
sched_item = kernel.schedule[sched_index]
if isinstance(sched_item, CallKernel):
assert not codegen_state.is_generating_device_code
from loopy.schedule import (gather_schedule_block,
get_insn_ids_for_block_at)
_, past_end_i = gather_schedule_block(kernel.schedule, sched_index)
assert past_end_i <= codegen_state.schedule_index_end
extra_args = synthesize_idis_for_extra_args(kernel, sched_index)
new_codegen_state = codegen_state.copy(
is_generating_device_code=True,
gen_program_name=sched_item.kernel_name,
schedule_index_end=past_end_i - 1,
implemented_data_info=(codegen_state.implemented_data_info +
extra_args))
from loopy.codegen.result import generate_host_or_device_program
codegen_result = generate_host_or_device_program(
new_codegen_state, sched_index)
glob_grid, loc_grid = kernel.get_grid_sizes_for_insn_ids_as_exprs(
get_insn_ids_for_block_at(kernel.schedule, sched_index))
return merge_codegen_results(codegen_state, [
codegen_result,
codegen_state.ast_builder.get_kernel_call(
codegen_state, sched_item.kernel_name, glob_grid, loc_grid,
extra_args),
])
elif isinstance(sched_item, EnterLoop):
tag = kernel.iname_to_tag.get(sched_item.iname)
from loopy.codegen.loop import (generate_unroll_loop,
generate_vectorize_loop,
generate_sequential_loop_dim_code)
from loopy.kernel.data import (UnrolledIlpTag, UnrollTag,
ForceSequentialTag, LoopedIlpTag,
VectorizeTag)
if isinstance(tag, (UnrollTag, UnrolledIlpTag)):
func = generate_unroll_loop
elif isinstance(tag, VectorizeTag):
func = generate_vectorize_loop
elif tag is None or isinstance(tag,
(LoopedIlpTag, ForceSequentialTag)):
func = generate_sequential_loop_dim_code
else:
raise RuntimeError("encountered (invalid) EnterLoop "
"for '%s', tagged '%s'" %
(sched_item.iname, tag))
return func(codegen_state, sched_index)
elif isinstance(sched_item, Barrier):
return codegen_state.ast_builder.emit_barrier(sched_item.kind,
sched_item.comment)
elif isinstance(sched_item, RunInstruction):
insn = kernel.id_to_insn[sched_item.insn_id]
from loopy.codegen.instruction import generate_instruction_code
return codegen_state.try_vectorized(
"instruction %s" % insn.id,
lambda inner_cgs: generate_instruction_code(inner_cgs, insn))
else:
raise RuntimeError("unexpected schedule item type: %s" %
type(sched_item))
def generate_code_v2(kernel):
"""
:returns: a :class:`CodeGenerationResult`
"""
from loopy.kernel import kernel_state
if kernel.state == kernel_state.INITIAL:
from loopy.preprocess import preprocess_kernel
kernel = preprocess_kernel(kernel)
if kernel.schedule is None:
from loopy.schedule import get_one_scheduled_kernel
kernel = get_one_scheduled_kernel(kernel)
if kernel.state != kernel_state.SCHEDULED:
raise LoopyError("cannot generate code for a kernel that has not been "
"scheduled")
# {{{ cache retrieval
from loopy import CACHING_ENABLED
if CACHING_ENABLED:
input_kernel = kernel
try:
result = code_gen_cache[input_kernel]
logger.debug("%s: code generation cache hit" % kernel.name)
return result
except KeyError:
pass
# }}}
from loopy.type_inference import infer_unknown_types
kernel = infer_unknown_types(kernel, expect_completion=True)
from loopy.check import pre_codegen_checks
pre_codegen_checks(kernel)
logger.info("%s: generate code: start" % kernel.name)
# {{{ examine arg list
from loopy.kernel.data import ValueArg
from loopy.kernel.array import ArrayBase
implemented_data_info = []
for arg in kernel.args:
is_written = arg.name in kernel.get_written_variables()
if isinstance(arg, ArrayBase):
implemented_data_info.extend(
arg.decl_info(
kernel.target,
is_written=is_written,
index_dtype=kernel.index_dtype))
elif isinstance(arg, ValueArg):
implemented_data_info.append(ImplementedDataInfo(
target=kernel.target,
name=arg.name,
dtype=arg.dtype,
arg_class=ValueArg,
is_written=is_written))
else:
raise ValueError("argument type not understood: '%s'" % type(arg))
allow_complex = False
for var in kernel.args + list(six.itervalues(kernel.temporary_variables)):
if var.dtype.involves_complex():
allow_complex = True
# }}}
seen_dtypes = set()
seen_functions = set()
seen_atomic_dtypes = set()
initial_implemented_domain = isl.BasicSet.from_params(kernel.assumptions)
codegen_state = CodeGenerationState(
kernel=kernel,
implemented_data_info=implemented_data_info,
implemented_domain=initial_implemented_domain,
implemented_predicates=frozenset(),
seen_dtypes=seen_dtypes,
seen_functions=seen_functions,
seen_atomic_dtypes=seen_atomic_dtypes,
var_subst_map={},
allow_complex=allow_complex,
var_name_generator=kernel.get_var_name_generator(),
is_generating_device_code=False,
gen_program_name=(
kernel.target.host_program_name_prefix
+ kernel.name
+ kernel.target.host_program_name_suffix),
schedule_index_end=len(kernel.schedule))
from loopy.codegen.result import generate_host_or_device_program
codegen_result = generate_host_or_device_program(
codegen_state,
schedule_index=0)
device_code_str = codegen_result.device_code()
from loopy.check import check_implemented_domains
assert check_implemented_domains(kernel, codegen_result.implemented_domains,
device_code_str)
# {{{ handle preambles
for arg in kernel.args:
seen_dtypes.add(arg.dtype)
for tv in six.itervalues(kernel.temporary_variables):
seen_dtypes.add(tv.dtype)
preambles = kernel.preambles[:]
preamble_info = PreambleInfo(
kernel=kernel,
seen_dtypes=seen_dtypes,
seen_functions=seen_functions,
# a set of LoopyTypes (!)
seen_atomic_dtypes=seen_atomic_dtypes)
preamble_generators = (kernel.preamble_generators
+ kernel.target.get_device_ast_builder().preamble_generators())
for prea_gen in preamble_generators:
preambles.extend(prea_gen(preamble_info))
codegen_result = codegen_result.copy(device_preambles=preambles)
# }}}
logger.info("%s: generate code: done" % kernel.name)
if CACHING_ENABLED:
code_gen_cache[input_kernel] = codegen_result
return codegen_result
def generate_code_for_sched_index(codegen_state, sched_index):
kernel = codegen_state.kernel
sched_item = kernel.schedule[sched_index]
if isinstance(sched_item, CallKernel):
assert not codegen_state.is_generating_device_code
from loopy.schedule import (gather_schedule_block,
get_insn_ids_for_block_at)
_, past_end_i = gather_schedule_block(kernel.schedule, sched_index)
assert past_end_i <= codegen_state.schedule_index_end
extra_args = synthesize_idis_for_extra_args(kernel, sched_index)
new_codegen_state = codegen_state.copy(
is_generating_device_code=True,
gen_program_name=sched_item.kernel_name,
schedule_index_end=past_end_i - 1,
implemented_data_info=(codegen_state.implemented_data_info +
extra_args))
from loopy.codegen.result import generate_host_or_device_program
codegen_result = generate_host_or_device_program(
new_codegen_state, sched_index)
glob_grid, loc_grid = kernel.get_grid_sizes_for_insn_ids_as_exprs(
get_insn_ids_for_block_at(kernel.schedule, sched_index))
return merge_codegen_results(codegen_state, [
codegen_result,
codegen_state.ast_builder.get_kernel_call(
codegen_state, sched_item.kernel_name, glob_grid, loc_grid,
extra_args),
])
elif isinstance(sched_item, EnterLoop):
tag = kernel.iname_to_tag.get(sched_item.iname)
from loopy.codegen.loop import (generate_unroll_loop,
generate_vectorize_loop,
generate_sequential_loop_dim_code)
from loopy.kernel.data import (UnrolledIlpTag, UnrollTag,
ForceSequentialTag, LoopedIlpTag,
VectorizeTag,
InOrderSequentialSequentialTag)
if isinstance(tag, (UnrollTag, UnrolledIlpTag)):
func = generate_unroll_loop
elif isinstance(tag, VectorizeTag):
func = generate_vectorize_loop
elif tag is None or isinstance(tag, (LoopedIlpTag, ForceSequentialTag,
InOrderSequentialSequentialTag)):
func = generate_sequential_loop_dim_code
else:
raise RuntimeError("encountered (invalid) EnterLoop "
"for '%s', tagged '%s'" %
(sched_item.iname, tag))
return func(codegen_state, sched_index)
elif isinstance(sched_item, Barrier):
# {{{ emit barrier code
from loopy.codegen.result import CodeGenerationResult
if codegen_state.is_generating_device_code:
barrier_ast = codegen_state.ast_builder.emit_barrier(
sched_item.synchronization_kind, sched_item.mem_kind,
sched_item.comment)
if sched_item.originating_insn_id:
return CodeGenerationResult.new(
codegen_state, sched_item.originating_insn_id, barrier_ast,
codegen_state.implemented_domain)
else:
return barrier_ast
else:
# host code
if sched_item.synchronization_kind in ["global", "local"]:
# host code is assumed globally and locally synchronous
return CodeGenerationResult(
host_program=None,
device_programs=[],
implemented_domains={},
implemented_data_info=codegen_state.implemented_data_info)
else:
raise LoopyError("do not know how to emit code for barrier "
"synchronization kind '%s'"
"in host code" %
sched_item.synchronization_kind)
# }}}
elif isinstance(sched_item, RunInstruction):
insn = kernel.id_to_insn[sched_item.insn_id]
from loopy.codegen.instruction import generate_instruction_code
return codegen_state.try_vectorized(
"instruction %s" % insn.id,
lambda inner_cgs: generate_instruction_code(inner_cgs, insn))
else:
raise RuntimeError("unexpected schedule item type: %s" %
type(sched_item))
def generate_code_v2(kernel):
"""
:returns: a :class:`CodeGenerationResult`
"""
from loopy.kernel import KernelState
if kernel.state == KernelState.INITIAL:
from loopy.preprocess import preprocess_kernel
kernel = preprocess_kernel(kernel)
if kernel.schedule is None:
from loopy.schedule import get_one_scheduled_kernel
kernel = get_one_scheduled_kernel(kernel)
if kernel.state != KernelState.SCHEDULED:
raise LoopyError("cannot generate code for a kernel that has not been "
"scheduled")
# {{{ cache retrieval
from loopy import CACHING_ENABLED
if CACHING_ENABLED:
input_kernel = kernel
try:
result = code_gen_cache[input_kernel]
logger.debug("%s: code generation cache hit" % kernel.name)
return result
except KeyError:
pass
# }}}
from loopy.type_inference import infer_unknown_types
kernel = infer_unknown_types(kernel, expect_completion=True)
from loopy.check import pre_codegen_checks
pre_codegen_checks(kernel)
logger.info("%s: generate code: start" % kernel.name)
# {{{ examine arg list
from loopy.kernel.data import ValueArg
from loopy.kernel.array import ArrayBase
implemented_data_info = []
for arg in kernel.args:
is_written = arg.name in kernel.get_written_variables()
if isinstance(arg, ArrayBase):
implemented_data_info.extend(
arg.decl_info(
kernel.target,
is_written=is_written,
index_dtype=kernel.index_dtype))
elif isinstance(arg, ValueArg):
implemented_data_info.append(ImplementedDataInfo(
target=kernel.target,
name=arg.name,
dtype=arg.dtype,
arg_class=ValueArg,
is_written=is_written))
else:
raise ValueError("argument type not understood: '%s'" % type(arg))
allow_complex = False
for var in kernel.args + list(six.itervalues(kernel.temporary_variables)):
if var.dtype.involves_complex():
allow_complex = True
# }}}
seen_dtypes = set()
seen_functions = set()
seen_atomic_dtypes = set()
initial_implemented_domain = isl.BasicSet.from_params(kernel.assumptions)
codegen_state = CodeGenerationState(
kernel=kernel,
implemented_data_info=implemented_data_info,
implemented_domain=initial_implemented_domain,
implemented_predicates=frozenset(),
seen_dtypes=seen_dtypes,
seen_functions=seen_functions,
seen_atomic_dtypes=seen_atomic_dtypes,
var_subst_map={},
allow_complex=allow_complex,
var_name_generator=kernel.get_var_name_generator(),
is_generating_device_code=False,
gen_program_name=(
kernel.target.host_program_name_prefix
+ kernel.name
+ kernel.target.host_program_name_suffix),
schedule_index_end=len(kernel.schedule))
from loopy.codegen.result import generate_host_or_device_program
codegen_result = generate_host_or_device_program(
codegen_state,
schedule_index=0)
device_code_str = codegen_result.device_code()
from loopy.check import check_implemented_domains
assert check_implemented_domains(kernel, codegen_result.implemented_domains,
device_code_str)
# {{{ handle preambles
for arg in kernel.args:
seen_dtypes.add(arg.dtype)
for tv in six.itervalues(kernel.temporary_variables):
seen_dtypes.add(tv.dtype)
preambles = kernel.preambles[:]
preamble_info = PreambleInfo(
kernel=kernel,
seen_dtypes=seen_dtypes,
seen_functions=seen_functions,
# a set of LoopyTypes (!)
seen_atomic_dtypes=seen_atomic_dtypes,
codegen_state=codegen_state
)
preamble_generators = (kernel.preamble_generators
+ kernel.target.get_device_ast_builder().preamble_generators())
for prea_gen in preamble_generators:
preambles.extend(prea_gen(preamble_info))
codegen_result = codegen_result.copy(device_preambles=preambles)
# }}}
# For faster unpickling in the common case when implemented_domains isn't needed.
from loopy.tools import LazilyUnpicklingDict
codegen_result = codegen_result.copy(
implemented_domains=LazilyUnpicklingDict(
codegen_result.implemented_domains))
logger.info("%s: generate code: done" % kernel.name)
if CACHING_ENABLED:
code_gen_cache.store_if_not_present(input_kernel, codegen_result)
return codegen_result
def generate_code_for_sched_index(codegen_state, sched_index):
kernel = codegen_state.kernel
sched_item = kernel.schedule[sched_index]
if isinstance(sched_item, CallKernel):
assert not codegen_state.is_generating_device_code
from loopy.schedule import (gather_schedule_block, get_insn_ids_for_block_at)
_, past_end_i = gather_schedule_block(kernel.schedule, sched_index)
assert past_end_i <= codegen_state.schedule_index_end
extra_args = synthesize_idis_for_extra_args(kernel, sched_index)
new_codegen_state = codegen_state.copy(
is_generating_device_code=True,
gen_program_name=sched_item.kernel_name,
schedule_index_end=past_end_i-1,
implemented_data_info=(codegen_state.implemented_data_info
+ extra_args))
from loopy.codegen.result import generate_host_or_device_program
codegen_result = generate_host_or_device_program(
new_codegen_state, sched_index)
glob_grid, loc_grid = kernel.get_grid_sizes_for_insn_ids_as_exprs(
get_insn_ids_for_block_at(kernel.schedule, sched_index))
return merge_codegen_results(codegen_state, [
codegen_result,
codegen_state.ast_builder.get_kernel_call(
codegen_state,
sched_item.kernel_name,
glob_grid, loc_grid,
extra_args),
])
elif isinstance(sched_item, EnterLoop):
tags = kernel.iname_tags(sched_item.iname)
tags = tuple(tag for tag in tags if tag)
from loopy.codegen.loop import (
generate_unroll_loop,
generate_vectorize_loop,
generate_sequential_loop_dim_code)
from loopy.kernel.data import (UnrolledIlpTag, UnrollTag,
ForceSequentialTag, LoopedIlpTag, VectorizeTag,
InOrderSequentialSequentialTag, filter_iname_tags_by_type)
if filter_iname_tags_by_type(tags, (UnrollTag, UnrolledIlpTag)):
func = generate_unroll_loop
elif filter_iname_tags_by_type(tags, VectorizeTag):
func = generate_vectorize_loop
elif not tags or filter_iname_tags_by_type(tags, (LoopedIlpTag,
ForceSequentialTag, InOrderSequentialSequentialTag)):
func = generate_sequential_loop_dim_code
else:
raise RuntimeError("encountered (invalid) EnterLoop "
"for '%s', tagged '%s'"
% (sched_item.iname, ", ".join(str(tag) for tag in tags)))
return func(codegen_state, sched_index)
elif isinstance(sched_item, Barrier):
# {{{ emit barrier code
from loopy.codegen.result import CodeGenerationResult
if codegen_state.is_generating_device_code:
barrier_ast = codegen_state.ast_builder.emit_barrier(
sched_item.synchronization_kind, sched_item.mem_kind,
sched_item.comment)
if sched_item.originating_insn_id:
return CodeGenerationResult.new(
codegen_state,
sched_item.originating_insn_id,
barrier_ast,
codegen_state.implemented_domain)
else:
return barrier_ast
else:
# host code
if sched_item.synchronization_kind in ["global", "local"]:
# host code is assumed globally and locally synchronous
return CodeGenerationResult(
host_program=None,
device_programs=[],
implemented_domains={},
implemented_data_info=codegen_state.implemented_data_info)
else:
raise LoopyError("do not know how to emit code for barrier "
"synchronization kind '%s'" "in host code"
% sched_item.synchronization_kind)
# }}}
elif isinstance(sched_item, RunInstruction):
insn = kernel.id_to_insn[sched_item.insn_id]
from loopy.codegen.instruction import generate_instruction_code
return codegen_state.try_vectorized(
"instruction %s" % insn.id,
lambda inner_cgs: generate_instruction_code(inner_cgs, insn))
else:
raise RuntimeError("unexpected schedule item type: %s"
% type(sched_item))
def generate_code_for_sched_index(codegen_state, sched_index):
kernel = codegen_state.kernel
sched_item = kernel.schedule[sched_index]
if isinstance(sched_item, CallKernel):
assert not codegen_state.is_generating_device_code
from loopy.schedule import (gather_schedule_block, get_insn_ids_for_block_at)
_, past_end_i = gather_schedule_block(kernel.schedule, sched_index)
assert past_end_i <= codegen_state.schedule_index_end
extra_args = synthesize_idis_for_extra_args(kernel, sched_index)
new_codegen_state = codegen_state.copy(
is_generating_device_code=True,
gen_program_name=sched_item.kernel_name,
schedule_index_end=past_end_i-1,
implemented_data_info=(codegen_state.implemented_data_info
+ extra_args))
from loopy.codegen.result import generate_host_or_device_program
codegen_result = generate_host_or_device_program(
new_codegen_state, sched_index)
glob_grid, loc_grid = kernel.get_grid_sizes_for_insn_ids_as_exprs(
get_insn_ids_for_block_at(kernel.schedule, sched_index))
return merge_codegen_results(codegen_state, [
codegen_result,
codegen_state.ast_builder.get_kernel_call(
codegen_state,
sched_item.kernel_name,
glob_grid, loc_grid,
extra_args),
])
elif isinstance(sched_item, EnterLoop):
tag = kernel.iname_to_tag.get(sched_item.iname)
from loopy.codegen.loop import (
generate_unroll_loop,
generate_vectorize_loop,
generate_sequential_loop_dim_code)
from loopy.kernel.data import (UnrolledIlpTag, UnrollTag, ForceSequentialTag,
LoopedIlpTag, VectorizeTag)
if isinstance(tag, (UnrollTag, UnrolledIlpTag)):
func = generate_unroll_loop
elif isinstance(tag, VectorizeTag):
func = generate_vectorize_loop
elif tag is None or isinstance(tag, (LoopedIlpTag, ForceSequentialTag)):
func = generate_sequential_loop_dim_code
else:
raise RuntimeError("encountered (invalid) EnterLoop "
"for '%s', tagged '%s'" % (sched_item.iname, tag))
return func(codegen_state, sched_index)
elif isinstance(sched_item, Barrier):
return codegen_state.ast_builder.emit_barrier(
sched_item.kind, sched_item.comment)
elif isinstance(sched_item, RunInstruction):
insn = kernel.id_to_insn[sched_item.insn_id]
from loopy.codegen.instruction import generate_instruction_code
return codegen_state.try_vectorized(
"instruction %s" % insn.id,
lambda inner_cgs: generate_instruction_code(inner_cgs, insn))
else:
raise RuntimeError("unexpected schedule item type: %s"
% type(sched_item))
